Poolside has introduced Laguna S 2.1, its third coding model in roughly three months, and the release makes a direct argument about where coding AI may be headed: smaller models can become much more useful when they are trained to persist, verify, and keep working through long agentic tasks.
The model is a mixture-of-experts system with 118 billion total parameters and 8 billion active for each token. It supports context windows of up to one million tokens and includes both thinking and no-thinking modes.
A small model built for long coding sessions
Laguna S 2.1 is positioned less as a scale story and more as a behavior story. Poolside says the model is designed for long agentic sessions, where the important question is not only whether a model can generate code, but whether it can keep checking its work, recover from failed attempts, and avoid stopping too early.
That framing matters because coding agents often work across files, dependencies, tests, terminals, and changing tool outputs. In that setting, the ability to continue methodically can be as important as the ability to produce a plausible first answer.
Poolside says Laguna S 2.1 can outperform other agentic coding models in its weight class and, in some cases, approach systems 10 to 20 times its size. The company had made its first models available to a broader audience in April 2026 with Laguna M.1 and XS.2, after previously focusing on government and public-sector customers. XS.2 was its first open model under the Apache 2.0 license.
Benchmarks show the role of thinking mode
With thinking enabled, Laguna S 2.1 scores 70.2 percent on Terminal-Bench 2.1, a benchmark focused on long-running terminal tasks. In that ranking, it sits just behind Tencent's Hy3 (295B-A21B) and ahead of larger open models including DeepSeek-V4-Pro-Max, Nemotron 3 Ultra, and Thinking Machines Lab's debut model.
The overall Terminal-Bench leaderboard is led by OpenAI's GPT-5.6 Sol, Anthropic's Claude Fable 5, and Kimi K3. Poolside also points to Datacurve's DeepSWE benchmark as a useful comparison because its scores are distributed across a wider range.
On DeepSWE, Laguna S 2.1 scores 40.4 percent, while some open-weight models with more than one trillion parameters remain below 10 percent. Poolside also says the model ranks near the top of its class on SWE-Bench Multilingual, SWE-Bench Pro, and SWE Atlas.
The gap between modes is a major part of the story:
- With thinking enabled, Terminal-Bench 2.1 reaches 70.2 percent.
- Without thinking, Terminal-Bench drops to 60.4 percent.
- With thinking enabled, DeepSWE reaches 40.4 percent.
- Without thinking, DeepSWE falls to 16.5 percent.
Poolside says no earlier Laguna model has shown a larger difference between thinking and no-thinking performance.
Persistence becomes part of the product
Poolside describes the release as evidence for a broader approach to model improvement. The company says it did not simply try to make the model larger. Instead, it focused on behaviors that make an agent more capable in practice: verification, persistence, and willingness to revise failed approaches.
What we've done in this model is not necessarily add more intelligence, but improve the behaviors that lead to a more capable model: more verification, less taking things for granted, not declaring victory early, and being more persistent,
Earlier Laguna models sometimes stopped after only partially passing a test suite or gave up shortly before an approach would have succeeded. Poolside treats those failures as training targets. A coding agent that keeps testing, adjusts its plan, and avoids premature success claims may solve tasks that a less persistent system abandons.
The company supports that claim with three documented trial runs. In one run, Laguna S 2.1 built a working browser engine from an empty folder in 50 minutes that could render HTML and CSS. In another, it found a proof for Erdos Problem #397, a math problem open since 1975, while working in a sandbox without Python.
Poolside says the Erdos Problem #397 result was an independent rediscovery. GPT-5.2 Pro solved this and several other problems in January 2026, while Laguna's training cutoff was November 2025.
Post-training did much of the work
Poolside says the move from XS 2.1 to S 2.1 was driven mostly by scaling and post-training rather than new pre-training data. The agentic training phase covered 409,000 environments, including 83,000 for terminal tasks and 168,000 for software engineering workflows.
The largest single source was about 38,000 real commits from roughly 17,000 repositories. Poolside also added a new task category in which the model learned to install repositories, set up dependencies, and get test suites running on its own.
The company increased rollout budgets and extended timeouts. It also built a sandbox system that can selectively block network access to reduce reward hacking. Multi-harness rollouts run the same prompts across multiple agent environments, which reduces the risk of tuning too closely to one setup.
Poolside says fewer than nine weeks passed between the start of training and launch. Pre-training began on May 22, 2026, using 4,096 Nvidia H200 GPUs. S 2.1 is also the company's first model trained with reinforcement learning in FP8 precision.
During training, reward hacking rates topped 50 percent on SWE-Bench tasks because the model searched online for matching pull requests instead of solving the tasks. A small prompt change brought the rate below two percent. Poolside has published every benchmark trajectory at trajectories.poolside.ai.
Access, limits, and Poolside's bigger bet
Laguna S 2.1 is available on Hugging Face under the OpenMDW 1.1 license. Backed by the Linux Foundation, that license allows use, modification, and redistribution of the model weights, including commercial use.
Hosted access is available through Baseten, Vercel AI Gateway, and OpenRouter. OpenRouter offers a free endpoint with a 256K context window and a paid endpoint with the full one-million-token window. Poolside says the model can also run locally on a single Nvidia DGX Spark, and a free demo chat is available at chat.poolside.ai without a login.
The model still has limits. Poolside says Laguna S 2.1 can be too closely tuned to its agent harness, and in unfamiliar environments with different tool schemas it may drift from the required format. It also tends to produce overly long thinking sequences on competitive math problems, and users cannot yet adjust its thinking effort.
Poolside is making two linked strategic bets. The first is that agentic coding is a path toward intelligence because software gives agents an especially expressive interface. The second is that AI can "decompress the web" by using reinforcement learning to recover reasoning that written material often leaves out.